Risky Behavior Via Social Media: The Role of Reasoned and Social Reactive Pathways

2 Objectives. It is important to understand what factors make some users of social media 3 engage in risky activities. This under-researched area is the focus of the present study which 4 applies the dual-process Prototype Willingness Model to demonstrate the potential role of 5 reasoned and social reactive pathways in explaining risk behaviors in adolescents and adults 6 in the online environment. Design. Quantitative single time point study using online survey 7 data from an international sample of social media users ( N = 1220). Methods. Two-step 8 logistic regression analysis tested the predictive ability of the reactive pathway variables of the Prototype Willingness Model above and beyond reasoned pathway variables from 10 expectancy-value models such as the Theory of Reasoned Action and Theory of Planned 11 Behavior. Results. The reactive pathway variables increased explained variance in 12 willingness to engage in online risk behaviors (compared to reasoned pathway variables 13 alone) by a mean improvement of 6.2% across both adolescent and adult age groups.


Introduction
27 Social media sites such as Facebook, Twitter, and YouTube offer opportunities for users to 28 interact and share information not only with their friends and family but also with people who 29 have similar interests. Over recent years the number of people using such sites has increased 30 dramatically (Perrin, 2015) and people of all ages are permanently logged onto social media 31 through their cell phones and mobile tablets (Peters & Allouch, 2005). However, alongside 32 the benefits such as improved socialization and communication and enhanced learning 33 opportunities, social media use can also pose specific risks such as cyberbullying, sexting, 34 sending embarrassing photos, publicly sharing location, and the spread of dangerous pranks 35 and games like the 'Choking Game' (Ahern, Sauer, & Thacker, 2015;Branley & Covey, 36 2017;GASP, 2013;Garner & O'Sullivan, 2010;O'Keeffe & Clarke-Pearson, 2011;Tsai, 37 Kelley, Cranor, & Sadeh, 2010).
38 It is important to understand which factors may influence some users to engage in these types 39 of risky social media practices. People might not be aware of the risks involved or they might 40 underplay the risks associated with social media use. They might also be subject to social 41 pressure and be influenced by whether the activity is commonplace amongst their peers. 42 However, little is known about the role of these or other types of social cognitive factors. To 43 fill this gap the present research adopted a dual-process framework of the type set out in the 44 Prototype Willingness Model (PWM: Gerrard, Gibbons, Houlihan, Stock, & Pomery, 2008;45 Gibbons, Gerrard, Blanton, & Russell, 1998) to predict willingness to engage in four different 46 types of risky online activities: sharing embarrassing photos, publicly sharing one's current 47 location, engaging in and sharing the videos of risky pranks and stunts, and engaging in 48 sexual communication with strangers. These four behaviors were chosen as we wished to 49 investigate risk taking behavior which reflects behaviors at the heart of social media: sharing, 50 i.e., location sharing, photo sharing and online communication; and these behaviors have 51 previously been linked to social media usage (Brake, 2014). 52 The reasoned pathway antecedents proposed in models like the Theory of Reasoned Action 53 (TRA: Fishbein & Ajzen, 1975), Theory of Planned Behavior (TPB: Ajzen, 1991) and 54 Fishbein's (2008) integrative model of behavioral prediction (IM) have been widely 55 successful in predicting positive health behaviors. However they have not been as 56 successfully applied to the prediction of negative or risky behaviors. It has been suggested 57 that this may be due to the models being focused purely upon a reasoned, intentional pathway 58 to risk. The PWM incorporates two different pathways to behavior: a reasoned pathway to 59 account for risk behaviors that are planned and determined by intentions, and a social reactive 60 pathway to account for unplanned or non-intentional variations in people's willingness to 61 engage in risk behavior.
62 Dual-process models, like the PWM, are based on the assumption that there are two types of 63 decision making involved in health behavior. The first type of decision making is analytical 64 and based upon the idea that behavior is planned and intentional. The PWM conceptualizes 65 this as a reasoned action pathway similar to that described in models such as the TRA 66 (Fishbein & Ajzen, 1975), TPB (Ajzen, 1991) and the IM (Fishbein, 2008). Antecedents of 67 this reasoned pathway which have been shown to account for a considerable proportion of the 68 variance in a range of health behaviors include people's attitudes towards the behavior (e.g., 69 whether the individual perceives the behavior as positive or negative) and their perceptions of 70 the social pressures to perform or not perform a behavior -which as outlined in the IM can 71 be a function of both injunctive norms (perceptions of whether the behavior is approved or 72 disapproved by others) and descriptive norms (whether others are engaging in the behavior). 4 120 impulsive behaviors -which applies to a lesser or greater extent across the four activities. For 121 example, sharing ones location or embarrassing photos on social media might be considered 122 less risky than engaging with sexual communications with a stranger or engaging in and 123 sharing videos of risky pranks and stunts. Comparison between age groups also enabled us to 124 examine the extent that reactive-based decision processes may be exclusive to adolescents or 125 whether they appear to continue into adulthood. 126 127 2 Method 128 2.1 Sample and survey methodology 129 A single time point online survey provided data from a diverse sample of 1102 international 130 social media users from 77 countries; with the majority of participants from the UK, Ireland, 131 USA and Canada (refer to Appendix A for complete demographics). Participants were aged 132 between 13 and 80 years (M = 28.5 years, SD = 11.3 years); 69.7% were female and 30.3% 133 were male. The bias towards female participants appears to be representative of social media 134 users (Kimbrough, et al., 2013). Although findings suggest that this gender difference is 135 diminishing (Perrin, 2015), excluding results from online forums, there still appears be more 136 females using many of the social media platforms (e.g., Duggan et al., 2014;Hargittai, 2007;137 Madden & Zickuhr, 2011). However, it is also possible that the greater amount of female 138 participants could -at least partially -be due to a gender difference in responding to 139 questionnaires (e.g., Hill, Roberts, Ewings, & Gunnell, 1997). Although there were more 140 females than males in the sample, males still accounted for more than 30% of the sample; 141 therefore this gender difference was not considered problematic. 142 143 The survey was designed by the authors, reviewed by an expert within the field of social 144 media research and received ethical approval from the Durham University ethics committee. 145 The survey was also piloted on a small sample of participants via opportunistic sampling and 146 feedback was obtained regarding the clarity of the survey items and any difficulties 147 encountered by the participants. The survey was revised following this feedback and all 148 necessary amendments were made and piloted prior to recruitment. To help maintain 149 participants interest and to encourage completion of the entire survey, interesting and/or 150 humorous facts were displayed throughout the survey (Branley, Covey, & Hardey, 2014). To 151 be eligible to participate, users were required to be fluent English speakers and to have 152 accessed social media at least once in the last 3-month period. Almost 75% of the sample 153 reported using social media more than several times per day (Appendix B). All minors (<16 154 years) were recruited through schools and parental and participant consent was obtained prior 155 to participation. Minors completed the survey outside of school time. Adults were recruited 156 online via a range of social media channels (see Appendix C). As compensation for their 157 time, all participants had the option to enter a free prize draw for a £50 Amazon voucher. 158 Within this sample there were some surveys with incomplete data. This missing data was 159 tested for randomness using Little's MCAR (Missing Completely At Random) test. The 160 results were non-significant indicating that the data was missing completely at random. 161 Consequently, the missing data were addressed using Maximum Likelihood Estimation 162 which has been shown to be a reliable method for dealing with missing data, superior to the 163 deletion of incomplete cases (Enders & Bandalos, 2001 253 by calculating the difference between the participants' own scores on the TIPI and the 254 scenario rated scores for each of the five personality traits: Extraversion, Agreeableness, 255 Conscientiousness, Emotional Stability/Neuroticism, and Openness. The five difference 256 scores were then summed to create an overall difference score. This was then deducted from 257 20 (the largest difference score possible) to reverse the scores into a similarity score, i.e., high 258 scores represent high similarity and low scores represent low similarity. 259 260 As Gibbons et al. (1998) suggest that the strength of prototypes will be greatest when users 261 perceive the prototype as similar and as favorable, the interaction between the two variables 262 is also included, i.e., prototype similarity x prototype favorability. 263 264 2.3 Analysis 265 Two-step logistic regression analysis was used to assess whether the reactive pathway 266 antecedents (i.e., prototype similarity, prototype favorability) enhanced the prediction of 267 willingness to engage in online risk, above and beyond the reasoned pathway components 268 (i.e., attitudes, injunctive norms, descriptive norms and previous behavior). The first step 269 therefore included past behavior, attitudes and injunctive norms and descriptive norms. The 270 second step introduced the prototype variables (prototype similarity and prototype 271 favorability). As Gibbons et al. (1998) suggest that the strength of prototypes will be greatest 272 when users perceive the prototype as similar and as favorable, the interaction between the 273 two variables was also included in the second step (prototype similarity x prototype 274 favorability). To compare the predictive ability of the reasoned and reactive components 275 between adolescents and adults the regressions were run separately for respondents aged 19 276 years or under (N=258) and respondents aged 20 years or over (N=844). Refer to Appendix A 277 for full sample demographics.

3 Results
279 Prior to running the regression analyses, descriptive statistics were computed to confirm that 280 there was adequate variance on the dependent variable and predictors for both age groups 281 (i.e., there was no evidence that participants were all selecting the same value on the scale, 282 such as floor or ceiling effects). The results shown in Table 1.   283  << INSERT TABLE 1 HERE >>> 284 Checking for multicollinearity also revealed no cause for concern, with most correlations 285 between the predictors < .4 (Table 2). Multicollinearity was also tested during the regression 286 analyses and all VIF values were low (<5).

<< INSERT TABLE 2 HERE >>>
288 As shown in Table 3 the regression showed that the variables entered at step 1 (attitudes, 289 injunctive norms, descriptive norms and previous behavior) were highly significant, positive 290 predictors of willingness across all four risk behavior scenarios. This applied to both age 291 groups. Overall these variables accounted for between 26.2 -53.1% of the variance in 292 willingness to engage in the risk behavior.
293 Introducing the prototype variables in step 2 resulted in a significant increase in explained 294 variance across almost all of the scenarios for both age groups, with the exception of scenario 295 1 where the increase did not reach significance for adolescents Overall, across all of the 8 296 scenarios, explained variance was increased slightly more in adolescents (4.6-13.7%) than it 297 was in adults (2-10.7%).
298 The overall model explained higher total variance in willingness [to engage in risk behavior] 299 for the adult age group compared to the adolescent age group, across 3 of the 4 scenarios 300 (sharing embarrassing photos, sharing location and sharing sexual content). However, the 301 majority of this difference is accounted for by the attitudes and norms variables that were 302 entered in the first step of the regressions. The difference in explained variance between the 303 first and second steps in the regression (i.e., as a consequence of the introduction of the 304 reactive prototype-based variables prototype similarity and prototype favorability) was 305 generally greater for the adolescent group (Table 3).

<< INSERT TABLE 3 HERE >>
307 Of the two prototype variables, prototype favorability emerged as the most consistent 308 predictor of willingness. Favorability was a significant predictor of willingness across both 309 age groups and all four behavior scenarios. In comparison, the significance of prototype 310 similarity (and the interaction between prototype favorability and similarity) differed 311 according to the risk behavior and age group involved (Table 3). In order to explore this 312 further, additional regression analyses were carried out for each of the personality traits 313 individually. This allowed us to investigate how each of the personality traits interact with 314 similarity ratings and their effect upon willingness to engage in risk behavior. Of the five 315 personality traits, conscientiousness was the only trait to be significant in at least one of the 316 two age groups, across all four scenarios. Extraversion was significant in at least one of the 317 age groups for three of the four scenarios. Suggesting that similarity on extraversion and 318 conscientiousness may play a greater role in willingness to engage in risk behavior, than 319 some of the other personality traits. This may be a direct or indirect effect (the latter via an 320 interaction with perceived favourability) dependent upon the risk behavior in question. For 321 example, when investigating willingness to share (or tolerate) embarrassing photos online, 322 individuals were more willing to tolerate such photos if they were similar on 323 conscientiousness, but only if they also judged the prototype favorably. Whereas for sharing 324 location publicly online, adults were more willing to do so if they perceived themselves to be 325 similar to the prototype for ratings of conscientiousness, regardless of whether they perceived 326 the prototype to be favorable or not. 327 The remaining personality traits (agreeableness, emotional stability and openness) did play a 328 role to a lesser degree. The results are shown in full in Table 4. 339 Of the variables unique to the PWM, the increase in explained variance in willingness 340 appears to be mainly due to the prototype favorability factor, i.e., how favorably individuals 341 judge others who engage in the specific risk behavior. This differs to findings by Rivis et al. 342 (2006) who found a similar increase in predictive ability for the PWM variables (in relation 343 to drinking behavior, unhealthy food consumption and smoking) but found prototype 344 similarity to be the more reliable predictor. It is possible that this is due to assessing the 345 PWM in relation to different risk behaviors, or due to Rivis et al. using intention as their 346 dependent variable rather than willingness. As the reactive pathway of the PWM is designed 347 to explain willingness this was chosen as the most appropriate outcome variable for the 348 current study. Todd, Kothe, Mullan, & Monds (2016) recent review suggests that prototype 349 favorability has a relationship on behavior through willingness whereas prototype similarity 350 appears to demonstrate a stronger relationship with intention rather than willingness. The 351 latter is unexpected as the PWM proposes that both prototype variables influence behavior 352 through willingness (which in turn can impact intention) and the model does not include a 353 pathway directly through intention. However, a direct pathway to intention may explain why 354 Rivis et al. found similarity to be the more significant predictor. Future research should seek 355 to determine which factors influence behavior through willingness and which may have a 356 more direct route via intention. It is acknowledged that future studies could benefit from the 357 inclusion of a measure of intention (in addition to measuring willingness) to allow full testing 358 of the PWM and comparison to other models such as the reasoned action approach. It is also 359 possible that some online risk behaviors may be more reasoned in nature than others, e.g., 360 sharing location online for perceived benefits of making location known to others. Therefore 361 intention may explain these behaviors more than willingness alone. It is also worth noting 362 that the current study used a novel measure of prototype similarity, which may also account 363 for some of the differences in the predictive ability of this factor compared to previous 364 studies. Whereas previous studies have generally relied upon self-reported impressions of 365 similarity (e.g., "In general, how similar are you to the type of person who drinks four units 366 of alcohol and drives thereafter?", Rivis et al., 2011), these measures may be prone to 367 response bias. Similarity is a relatively abstract concept therefore the current study aimed to 368 include a potentially more objective measure of similarity by asking participants to rate the 369 prototypes on personality trait measures (using the TIPI). These measures were then 370 compared to their own personality trait measures to create a similarity/difference score. As no 371 statements about similarity or comparisons were provided to the participants, this method 372 may be less likely to introduce response bias. However it is possible that the current study 373 and previous research measures of similarity are tapping into slightly different constructs. 374 Interestingly, the inclusion of personality traits as a measure of similarity allowed us to run 375 further analyses to investigate whether some personality traits play a stronger role [compared 376 to others] in relation to willingness to engage in risk. The results suggest that similarity on 377 conscientiousness and extraversion may influence willingness to engage in online risk to a 378 greater degree than the other personality traits. The results also indicate that the predictive 379 ability of specific traits varies according to the risk behavior involved. Further research may 380 wish to investigate this in more detail. 381 382 Descriptive norms were found to be a weak predictor of willingness with the exception of one 383 of the scenarios which depicted engaging in dangerous pranks and sharing the videos online. 384 This may suggest that the role of descriptive norms as a predictor of willingness differs 385 according to the risk behavior in question. For example, this scenario depicted a potentially 386 more obvious physical risk (e.g., balancing on high objects, lying in the middle of the road) 387 compared to the other scenarios (e.g., sharing location online, sharing sexual content, or 388 sharing embarrassing photos) where the risk may be less immediately apparent and/or of a 10 389 potentially less physical nature. Alternatively it is possible that descriptive norms have more 390 of an effect upon behavior through intention rather than willingness. This is another potential 391 avenue for future research incorporating an intention and willingness measure. Future 392 research may also wish to include a wider range of online behaviors and predictors, to 393 identify if the type of behavior and/or nature of the associated risk impacts upon the 394 predictive value of each of the variables.

<< INSERT
395 The second aim of this research was to investigate the predictive ability of the reactive 396 pathway to willingness to engage in risk for adolescents and adults. The overall model 397 (including both the reasoned and reactive variables) explained more variance in willingness 398 for the adult age group. A finding that may initially seem surprising considering the PWM 399 was designed to explain risk behavior in adolescents Gibbons et al., 400 1998). However, further investigation shows that the higher percentage of explained variance 401 in willingness in adults is accounted for by the reasoned variables -attitudes and injunctive 402 norms in particular. The addition of the reactive prototype-based variables (prototype 403 similarity and prototype favorability) actually showed a greater increase in explained 404 variance in willingness for the adolescent group. This is an important finding because it 405 suggests that factors relating to the more rational pathway may play a greater role in adults 406 willingness to engage in risk; supporting Gerrard and colleague's (2008) theory that 407 adolescents' greater willingness to engage in risk behavior is due to decision-making shifting 408 to a more reasoned, analytical process with age. That said, the social reactive variables still 409 significantly increased explained variance in adult willingness to engage in behavior, above 410 and beyond that explained by the reasoned action variables based purely upon rational 411 decision-making pathways; suggesting that reactive pathways to risk may still play a role in 412 adult social media users' willingness to engage in risk taking in the online environment 413 (albeit to a lesser extent than adolescent users). It is important to note that the current study is 414 based upon single time point survey data and does not include a measure of actual risk 415 behavior. In order to further investigate the role of reactive processes and willingness to 416 engage in risk behavior, future research should include a measure of subsequent behavior.
417 Due to space and time constraints and a desire to limit participant dropout rates, single item 418 measures were used in the current study. Whilst there may be advantages using multi-item 419 measures, the use of single item measures was not deemed problematic as many previous 420 studies investigating the PWM have applied such measures (e.g., willingness: Pomery et al. 421 2009;favorability: Rivis et al. 2011;2006;norms: Rivis et al., 2011. It has also been 422 demonstrated that single item measures can be sufficient for constructs that are "easily and 423 uniformly imagined" and in many instances more items can provide little additional 424 information, with one or two clear measures being able to outperform some scales with 425 multiple items (Bergkvist & Rossiter, 2007;Drolet & Morrison, 2001). 426 We acknowledge that the specific wording and details of the 'risk scenarios' provided in the 427 survey may have influenced respondents' responses. However this does not undermine the 428 internal validity of the current study because we were interested in whether the respondents' 429 perceptions of likeability and similarity relate to their willingness to engage in a similar 430 activity. However, future research wishing to draw further conclusions about the general 431 factors underlying such behaviors should seek to ensure neutrality of the wording used within 432 the scenarios. Also, it may be worth clarifying the audience more specifically in future 433 research as users may imagine different social media platforms when answering the items 434 about the hypothetical scenarios. Although the current study did specify that the scenarios 435 depicted information that users were sharing openly/publicly, the specific platform may still  Step 2 Step 1 Step 2 Step 1 Step 2 Step 1 Step 2 Step 2 Step 1 Step 2 Step 1 Step 2 Step 1 Step 2  Table 4. Standardized coefficients for the two-step regression analyses testing each personality trait individually. Adolescents 13-19 years (n = 258), Adults ≥ 20 years (n = 844).

Continued on next page
Step 1 Step 2